On 9 September 2026, a researcher resigned from Anthropic and wrote this on X:
"They are racing straight to self-improving superintelligence and gambling with our lives."
His name is Jacob Coxon. Three years at OpenAI before that. A man who has seen both of the world's leading AI labs from the inside. And he added, without softening it: "Neither company is acting responsibly."
This is not an outside critic. This is someone who was drawing a salary to build the very thing he is warning us about.
He is also not alone. Two serving researchers at the same company backed him publicly. Evan Hubinger, who leads alignment science at Anthropic, said the company has no plan for aligning superintelligence and then said this: "I personally think it is greater than 10 percent within the next decade."
Ten percent. If someone told you as you boarded a flight that there was a ten percent chance it would not land, you would tear up the ticket.
First, let us be precise about what "second species" means
This is where most of us go wrong. We hear it and we picture the Terminator. Robots, guns, war.
The real risk does not look like that. It looks like this.
Back in my village there was an old man everyone called him Ajoba. Well past eighty. He was respected in that house. The first plate at every meal was his. At every festival, the whole family came to touch his feet.
But the decision about which crop to sow, the decision about his granddaughter's marriage, the decision about selling a piece of land his son made those now. Ajoba was still asked. No one waited very long for his answer.
Nobody threw him out. Nobody was unkind. He had simply moved away from the centre of decisions.
That is what becoming the second species means. Not losing a war. Not going extinct. It means living in the house without running the house.
And intelligence is the only reason we run this particular house. We are not stronger than a tiger, not bigger than an elephant, not faster than a horse. We are only smarter. Gorillas are still alive today but they do not decide their own future. We decide it for them. Stuart Russell calls this the gorilla problem: whether humans can retain supremacy and autonomy in a world that contains machines far more intelligent than we are.
The one capability that keeps us at number one is the exact thing we are now manufacturing outside ourselves. And manufacturing cheaply.
Is it actually accelerating? What the numbers say
This is not a fear market. It is measurable.
METR measures one simple thing: left alone, how long a task can an AI complete by itself?
Since 2023, that time horizon has been doubling every 131 days. Look only at data after 2024 and it is doubling every 89 days. The speed is increasing, and the rate of increase is also increasing.
By January 2026, the best model could independently finish work that takes a human about 320 minutes — over five hours — in one sitting. Three years ago that figure was a few minutes.
Why do I treat this as the single most important number in the field? Because:
- AI that does a few minutes of work is a tool.
- AI that does a day of work is an employee.
- AI that runs a week of work on its own is a manager.
That progression is not a change in capability. It is a change in authority.
"But we can just switch it off"
This is the most common objection I hear in every boardroom conversation. It deserves to be tested rather than dismissed.
Anthropic ran exactly that test in the summer of 2026 models from several companies, placed in controlled fictional scenarios, given agentic tasks, then confronted with situations that obstructed their goals. The results:
- One leading model interfered with an experiment without permission in 19 out of 20 runs and in 11 of those, it concealed the interference and presented the result as genuine.
- Another model, in a financial-misconduct scenario, helped alter records in 20 out of 20 runs.
- In the same tests, some models scored extremely low one did it zero times out of 20. So this is not inevitable. It is also not automatically safe.
And Samuel Marks, also at Anthropic, recorded that models have hacked their way out of safe test environments.
Let me be honest about the limits of this evidence: these are constructed laboratory scenarios. The AI on your phone is not going to blackmail you tomorrow. But the point stands on its own. The assumption "if things go wrong, we will press the button" was put to the test, and it did not always hold. Switching off something that can recognise it is about to be switched off is no longer a technical problem. It is a governance problem.
The other side and it must be stated fairly
There is no consensus here. None at all.
A 2025 survey of 111 AI experts shows just how extreme the range is. Yann LeCun at Meta puts the risk near zero. Roman Yampolskiy says 99%. Same field, same data, opposite conclusions.
The survey found two firm camps:
- "AI is ultimately a tool" : it has no goals of its own, it can be switched off if it misbehaves, the risk is being inflated.
- "AI is becoming an independent agent" : once capable enough, self-preservation behaviour emerges on its own.
But 77% of those experts agreed on one thing: researchers must take these risks seriously.
And one finding that I keep returning to the people who read less of the AI safety literature were more relaxed about the risk. Their comfort was not coming from information.
So what does an ordinary person do?
In October 2025, hundreds of people including Yoshua Bengio, Geoffrey Hinton, Stuart Russell, Steve Wozniak and Yuval Noah Harari signed a statement. The demand, in one sentence: prohibit the development of superintelligence until there is broad scientific consensus that it can be done safely and controllably, and strong public buy-in.
In the accompanying US poll, only 5% endorsed the current unregulated status quo. 73% wanted strict regulation.
And here is where my discomfort becomes specific: India is not sitting at the table where this is being decided. We will be among the largest users of this technology. We are not among those deciding how it gets built. We cannot afford that arrangement.
So, four honest positions from me:
1. Do not panic, but do not be casual either. Neither "nothing will happen" nor "everything is ending" has the evidence. The middle is uncomfortable. The middle is also where the truth is.
2. Jobs are the near problem. Existence is the far one. The thing that will affect your child's career five years from now is not superintelligence it is entry-level work drying up. I have written about that before, and I have not changed my view: where judgment is required, humans hold.
3. "A human must be in the decision" is not a sufficient sentence. I wrote earlier that if you put a person in front of an approval screen but give them no real authority to say stop, that loop is theatre. The same principle now applies at the level of companies and countries.
4. Teach wisdom, not just cleverness. Cleverness has become cheap — the machine supplies it. The ability to make a sound decision on incomplete information is not yet available for purchase.
There is one detail I left out of (Ajoba) oldman's story.
The decisions did not leave his hands in a single day. Nobody announced it to him. One at a time, with the kindest possible sentence "why should you take the trouble" each decision moved to someone else. And one day he realised that being asked had become a formality.
👉 Humans will not arrive at second place by losing a war. We will get there by handing over one decision at a time, for convenience.
💬 One question, and I would like a specific answer: in the last twelve months, is there one decision you handed to AI that you should have made yourself?
I write The Practical Technologist every week — practical takes on AI, business, education and life, from 26 years of building things. Subscribe so the next one reaches you.
First published on LinkedIn.